Enhanced defect detection on steel surfaces using integrated residual refinement module with synthetic data augmentation

Abstract

Ensuring high-quality production in the steel manufacturing industry is crucial for efficiency, waste reduction, and cost minimization. Traditional manual inspection methods are often inconsistent, time-consuming, and prone to human error, making automated visual inspection essential for reliable quality control. Steel surface defect detection plays a critical role in identifying issues such as cracks, scratches, and corrosion, which can compromise product durability and performance. This study proposes a new deep learning-based defect segmentation model to enhance the accuracy and efficiency of steel defect detection. The model incorporates ResNet50, Residual Block (RB), Residual Squeeze-and-Excitation Block (RSB), and Residual Refinement Module (RRM) to improve deep feature extraction and segmentation precision. Extensive evaluations demonstrate that the proposed model achieves an impressive 87.8% mean Intersection over Union (mIoU), outperforming existing segmentation models. A custom dataset was created using a real production line image acquisition system, ensuring diverse defect representation. Additionally, Synthetic Defect Generation (SDG) techniques were applied to enhance the dataset and improve model robustness. The proposed model offers a scalable and automated defect detection solution, significantly improving quality control, reducing inspection time, and ensuring higher reliability in industrial applications.